Optimization method for unit selection speech synthesis based on synthesis quality predictions

The hidden Markov model(HMM) based unit selection speech synthesis method improves the automation of system construction and the stability of synthetic speech compared with the conventional unit selection synthesis method.However,the weights used to combine the different acoustic statistical models in this method cannot be obtained through automatic training and are difficult to tune manually.This paper presents an approach based on synthesis quality predictions to optimize these model weights.First,subjective evaluation results are used to develop a prediction model based on multivariate adaptive regression splines to predict the quality of synthetic speech using different model weights.Second,a pattern search algorithm is used to automatically search for the optimal weight based on the trained prediction model.Tests indicate that this method effectively optimize the model weights to improve the natural flow of synthetic speech.